GLASS: A Generative Recommender for Long-sequence Modeling via SID-Tier and Semantic Search

TL;DR

This work introduces GLASS, a generative recommender that injects long-term user interests into the retrieval stage by leveraging Semantic IDs (SIDs). It proposes SID-Tier to improve the initial SID $sid_1$ and semantic hard search to recalibrate subsequent tokens using retrieved long-term context, augmented by sparsity-aware strategies to combat data sparsity. Empirical results on TAOBAO-MM and KuaiRec demonstrate consistent outperformance of state-of-the-art baselines, with notable gains in early and mid-range ranking metrics. The approach enables more accurate long-sequence modeling in generative rec systems, offering practical benefits for large-scale recommendation tasks.

Abstract

Leveraging long-term user behavioral patterns is a key trajectory for enhancing the accuracy of modern recommender systems. While generative recommender systems have emerged as a transformative paradigm, they face hurdles in effectively modeling extensive historical sequences. To address this challenge, we propose GLASS, a novel framework that integrates long-term user interests into the generative process via SID-Tier and Semantic Search. We first introduce SID-Tier, a module that maps long-term interactions into a unified interest vector to enhance the prediction of the initial SID token. Unlike traditional retrieval models that struggle with massive item spaces, SID-Tier leverages the compact nature of the semantic codebook to incorporate cross features between the user's long-term history and candidate semantic codes. Furthermore, we present semantic hard search, which utilizes generated coarse-grained semantic ID as dynamic keys to extract relevant historical behaviors, which are then fused via an adaptive gated fusion module to recalibrate the trajectory of subsequent fine-grained tokens. To address the inherent data sparsity in semantic hard search, we propose two strategies: semantic neighbor augmentation and codebook resizing. Extensive experiments on two large-scale real-world datasets, TAOBAO-MM and KuaiRec, demonstrate that GLASS outperforms state-of-the-art baselines, achieving significant gains in recommendation quality. Our codes are made publicly available to facilitate further research in generative recommendation.

Figures (3)

• Figure 1: Overview of the proposed GLASS framework. (a) GLASS adopts the encoder-decoder architecture. The encoder processes short-term item sequences and SID-Tier to capture user interests. The decoder integrates encoder output and Semantic Search via dual cross attention. (b) SID-Tier aggregates long-term history into a unified interest vector to enhance the first codeword ($sid_1$) prediction. (c) Semantic Hard Search performs hard search based on the generated $sid_1$, followed by a gate to enhance the generation of fine-grained tokens ($sid_2, sid_3$). sparsity-aware augmentation ensures robust context by incorporating semantic neighbors or resizing the codebook.
• Figure 2: Performance comparison of SIDTier and SHS modules.
• Figure 4: Average Gate Value vs. Retrieved Sequence Length during Inference. The curves represent two codebook configurations: $[128, 128, 128]$ (left) and $[64, 128, 128]$ (right).